NVIDIA launched the DGX Spark as "an AI supercomputer on your desk": a golden box priced at US$ 4,000 (around R$ 22,000) with 128 GB of memory, capable of loading AI models that do not fit in almost any regular computer. Then the independent tests came in and the results were uncomfortable: in response speed, the NVIDIA box ties with a Mac mini that costs one third of the price, loses to Macs with the M5 Max chip, and gets beaten by a gaming graphics card, the RTX 5090. It is not a manufacturing defect. It is a number on the spec sheet that the marketing does not highlight: memory bandwidth.
This post explains, without technical jargon, why this happens, shows the numbers from public tests, and draws the lesson that matters for anyone thinking about running AI inside their own company: the right hardware is not the one with the most famous brand or the biggest memory. It is the one sized for your specific case.
Large memory says what fits. Bandwidth says what moves.
Think of a warehouse. The size of the warehouse (memory, measured in GB) defines how much inventory you can store. The width of the door (bandwidth, measured in GB per second) defines how much inventory moves in and out per minute. A massive warehouse with a narrow garage door stores a lot and moves very little.
With AI it works the same way, with one detail that almost no one mentions: when an AI writes a response, it needs to reread the entire model from memory for every single word it generates. It is not a one-time read at the start. It is the entire model, again, for every word. So the maximum writing speed is a simple calculation: bandwidth divided by model size.
The DGX Spark moves 273 GB per second. A model with 70 billion parameters, compressed in the format everyone uses, takes up about 40 GB. Divide one by the other: in the best theoretical scenario, 7 words per second. In practice, less. A person reads between 4 and 5 words per second, meaning the US$ 4,000 machine delivers text at almost exactly the speed at which you read it. For a single-user chat, that is manageable. For an operation handling dozens of simultaneous requests, document analyses, and agents working in batch, that is a line at the warehouse door.
The numbers, side by side
Memory bandwidth for each machine, according to official spec sheets:
- NVIDIA DGX Spark: 128 GB of memory, 273 GB/s, US$ 4,000.
- Mac mini (M4 Pro chip): up to 64 GB of memory, the same 273 GB/s, at about one third of the price.
- MacBook Pro with M5 Max chip: 614 GB/s, more than double the Spark.
- RTX 5090 (gaming graphics card): 32 GB of memory, 1,792 GB/s, six and a half times the Spark's bandwidth.
And real-world tests confirm the math. In the public llama.cpp benchmark (the most widely used tool in the world for running AI on your own computer), the DGX Spark generates around 35 words per second on a large 120 billion-parameter model. A professional NVIDIA GPU with high memory bandwidth does 180 words per second on the same task: five times faster. And medium-sized models that fit in the RTX 5090's 32 GB run on that gaming card many times faster than on the Spark, because its bandwidth is six times higher.
One important clarification: the Spark is not slow at everything. It reads questions and input documents very fast, because that depends on compute power, which it has in abundance. The bottleneck shows up when writing the response, which is exactly the part the user is waiting to see on screen.
"But it runs models the others cannot"
True. The 128 GB of shared memory loads models that do not fit in either the Mac mini or the RTX 5090. That is the entire sales argument. The problem is the word "runs." Loading a massive model and generating 3 to 7 words per second is like hiring a brilliant consultant who dictates their report at one word per second: the knowledge is there, but the result does not arrive in useful time.
Large shared memory without bandwidth is the giant warehouse with the narrow door. Good for storage, not for operations. That is why all independent tests reached the same conclusion: the Spark is an excellent lab machine for developers in the NVIDIA ecosystem to test large models before sending them to the data center. As a production machine for sustaining the daily operations of a business, the cost-per-speed ratio does not add up.
The lesson for companies that want AI in-house
This episode teaches more about buying technology than it does about NVIDIA. The mistake it exposes is classic: choosing AI hardware by logo or by the biggest number in the ad. The 128 GB of memory is the showroom number. The 273 GB/s bandwidth is the number that runs the day-to-day. Whoever does not know the difference pays R$ 22,000 for an experience slower than an entry-level Mac.
And this decision is landing on more and more company desks, because running AI in-house (what is called on-premise) solves two problems that the cloud cannot: business data stops traveling to third-party servers, and the monthly usage bill becomes a fixed, predictable cost. We already covered how to have AI without leaking company data and why it matters.
But in-house AI only pays off when the sizing is done in the right order: first the task that needs to be solved, then the smallest model that solves it with quality, then how many people and processes will use it simultaneously, and only then the machine that can handle all of that with room to spare. Sometimes the answer is a desktop Mac. Sometimes it is a gaming card in a simple server. Sometimes it is a server with professional GPUs. What the answer never is: "the most expensive box from the most famous brand."
That is exactly the sizing work that Steply does in the on-premise AI diagnostic: we assess your case, run the bandwidth, model, and concurrency calculation, and deliver the right machine specification before you spend a single real on hardware.
The NVIDIA golden box is not a bad product. It is a lab product being sold as a desktop solution, and whoever buys it without doing the math finds out the difference after receiving the bill. Spec sheets have showroom numbers and numbers that run the show. Knowing how to tell them apart is the difference between investing in AI and collecting equipment.